| import random |
| from training.stereo_backbones_v1.common import batch,waveform |
| from training.hvs_audio_fourtask.data import load_image |
|
|
| def pack(actor,msgs,images,answer,path): |
| import torch |
| if path:msgs[-1]['content'].append(dict(type='audio')) |
| audio=waveform(dict(audio_path=path),'vision_stereo') if path else None |
| b=actor.prepare_inputs(msgs,images,audio);n=b['input_ids'].shape[1] |
| eos=actor.proc.tokenizer.eos_token_id;assert eos is not None |
| ids=actor.proc.tokenizer.encode(answer,add_special_tokens=False)+[eos] |
| b['input_ids']=torch.cat([b['input_ids'],torch.tensor([ids])],1);b['attention_mask']=torch.ones_like(b['input_ids']) |
| for k in ['mm_token_type_ids','token_type_ids']: |
| if k in b:b[k]=torch.cat([b[k],torch.zeros((1,len(ids)),dtype=b[k].dtype)],1) |
| b['labels']=b['input_ids'].clone();b['labels'][:,:n]=-100 |
| assert b['input_ids'].shape[1]<=4096 |
| return b |
|
|
| def inverse(actor,row,seed): |
| rng=random.Random(seed);before=row['pose'];dy=rng.choice([-45,-30,-15,15,30,45]);dp=rng.choice([-15,0,15]);after=[(before[0]+dy)%360,max(-75,min(75,before[1]+dp))];dp=after[1]-before[1] |
| msgs=[dict(role='system',content='Infer the camera rotation between two views. The microphones are fixed and do not rotate with the camera.'),dict(role='user',content=[dict(type='text',text='Current view:'),dict(type='image'),dict(type='text',text='Next view:'),dict(type='image'),dict(type='text',text='Return one <answer>rotate(dyaw,dpitch)</answer>.')])] |
| pics=[load_image(dict(panorama=row['observations'][-1]['panorama'],pose=p)) for p in [before,after]] |
| return pack(actor,msgs,pics,f'<answer>rotate({dy},{dp})</answer>',row.get('audio_path')) |
|
|
| def conditional(actor,row,seed): |
| rng=random.Random(seed);before=[rng.choice(list(range(0,360,30))),rng.choice([-15,0,15])];indices=[0,1];rng.shuffle(indices) |
| pics=[load_image(dict(panorama=row['panorama'],pose=before))];content=[dict(type='text',text='Current view:'),dict(type='image')] |
| for j,i in enumerate(indices): |
| content.extend([dict(type='text',text=f'Candidate future view {j+1}:'),dict(type='image')]);pics.append(load_image(dict(panorama=row['panorama'],pose=row['candidates'][i]['center']))) |
| question={'first':'started first','last':'started later','longer':'lasted longer','shorter':'lasted for less time'}[row['query']] |
| content.append(dict(type='text',text=f'Use the recording to select the object whose sound {question}. Infer the camera rotation from the current view to that candidate view. Return one <answer>rotate(dyaw,dpitch)</answer>.')) |
| target=row['candidates'][row['target_index']]['center'];dy=round((target[0]-before[0]+180)%360-180);dp=round(target[1]-before[1]) |
| return pack(actor,[dict(role='system',content='Ground audible events in candidate future views. Microphones remain fixed.'),dict(role='user',content=content)],pics,f'<answer>rotate({dy},{dp})</answer>',row['audio_path']) |
|
|
